Deep Learning segmentation with metal intrusion for quantitative microstructure analysis of hardened cement paste

Quantitative microstructure analysis of hardened cement paste is critical but challenging due to the highly disordered nature of the materials. Here, we introduced an advanced approach combining metal intrusion and deep learning segmentation to enhance the microstructure analysis of the composites. First, a low-melting-point metal, Field metal, was injected into the cement samples under pressure in a vessel to enhance discrimination between pores and solid phases under backscatter electron (BSE) imaging. The subsequent segmentation of pore, unhydrated, and hydrated cement phases in BSE images of these metal-intruded hardened cement paste samples was performed using deep learning segmentation models such as Linknet and Unet. Microstructure parameter size and shape analysis were conducted for each segmented phase, including area, equivalent diameter, solidity, circularity, and aspect ratio. The results showed that segmenting the different phases in the hardened cement paste was possible using this approach. While both U-Net and LinkNet networks showed good segmentation results with mean IoU scores of 0.89 and 0.87, respectively, the U-Net outperformed with more details and complex boundaries. This segmentation simplifies the microstructure of cement composites, enabling the quantitative analysis of each phase. These detailed analyses are crucial for evaluating material behaviour and optimising the performance of cement composites.